Strongly Exponential Symmetric Spaces

STRONGLY EXPONENTIAL SYMMETRIC SPACES

YANNICK VOGLAIRE

Abstract. We study the exponential map of connected symmetric spaces and characterize, in terms of midpoints and of inﬁnitesimal conditions, when it is a diﬀeomorphism, generalizing the Dixmier–Saito theorem for solvable Lie groups. We then give a geometric characterization of the (strongly) exponential solvable symmetric spaces as those spaces for which every triangle admits a unique double triangle. This work is motivated by Weinstein’s quantization by groupoids program applied to symmetric spaces.

1. Introduction

In the late 1950’s, Dixmier [5] and Saito [16] characterized the strongly exponential Lie groups: the simply connected solvable Lie groups for which the exponential map is a global diﬀeomorphism. They are characterized by the fact that the exponential map is either injective or surjective, or equivalently the fact that the adjoint action of their Lie algebra has no purely imaginary eigenvalue. In contrast, the exponential map of semisimple Lie groups is never injective, but may be surjective or have dense image. As a result, the latter two weaker notions of exponentiality have been studied for general Lie groups by a number of authors (see [6, 21] for two surveys and for references). Recently, a number of their results were generalized to symmetric spaces in [15]. The ﬁrst main result of the present paper intends to complete the picture by giving a version of the Dixmier–Saito theorem for symmetric spaces (see Theorems 1.1, 1.2 and 1.3). The second main result gives a geometric characterization of the strongly ex- arXiv:1303.5925v2 [math.DG] 6 Apr 2014 ponential symmetric spaces which are homogeneous spaces of solvable Lie groups, in terms of existence and uniqueness of double triangles (see Theorem 1.4). This

Part of this work was done during the author’s PhD thesis at the Université Catholique de Louvain, Belgium, and the Université de Reims, France, under the supervision of Pierre Bieliavsky and Michael Pevzner. The author thanks Mathieu Carette for discussions on midpoints and geodesics in symmetric spaces. This work was partially supported for travel by the Belgian Science Policy (belspo) through the Interuniversity Attraction Pole (IAP) PHASE VII/18 project “DYGEST”. The published version of this paper is available at http://imrn.oxfordjournals.org/cgi/ content/abstract/rnt149?ijkey=My6gILrUCEG3Z1z&keytype=ref. 1 2 YANNICK VOGLAIRE is related to a beautiful theorem of A. Weinstein [20] relating the graph of multiplication in the pair groupoid of a symplectic symmetric space to the symplectic area of such double triangles. Weinstein’s theorem was developed in the context of his quantization by groupoids program and forms, together with the extension of Bieliavsky’s quantization of symplectic symmetric spaces [3], the initial motivation for the present work. Since we will only be concerned with strong exponentiality in this paper, from now on we will drop the adjective “strong”, although nowadays exponentiality has a weaker meaning in the literature.

A symmetric space is a manifold M endowed with a family of involutions {sx}x∈M of M, called symmetries, such that x is an isolated (not necessarily unique) ﬁxed point of sx, and such that sx(syz) = s(sxy)(sxz) for all x, y, z ∈ M.A morphism of symmetric spaces M,N is a smooth map φ : M → N such that φ(sxy) = sφ(x)φ(y) for all x, y ∈ M. Some basic examples to keep in mind are the following: n (Ex1) the n-dimensional Euclidean space R with symmetries sxy = 2x − y, (Ex2) the n-sphere Sn where the symmetry at x is the restriction to Sn of the axial symmetry about 0x,

(Ex3) the hyperbolic plane Π2 with the Riemannian geodesic symmetries, −1 (Ex4) any Lie group G with symmetries sgl = gl g, as well as two solvable two-dimensional examples: 2 0 0 0 0 0 (Ex5) the space A2 = R with symmetries s(a,b)(a , b ) = (2a−a , 2 cosh(a−a )b−b ), 2 0 0 0 0 0 (Ex6) the space B2 = R with symmetries s(a,b)(a , b ) = (2a−a , 2 cos(a−a )b−b ). The connected symmetric spaces in this sense (due to Loos [9]) coincide with the homogeneous symmetric spaces: the homogeneous spaces G/H of connected Lie groups G, with H an open subgroup of the fixed points set Gσ of an involutive automorphism σ of G (called an involution in what follows). The symmetries in −1 that case are given by sgH lH = gσ(g l)H. Such a triple (G, H, σ) will be called a symmetric triple. When H is the full fixed point subgroup Gσ, we simply refer to the symmetric pair (G, σ). A symmetric triple (G, H, σ) is said to realize a symmetric space M if G/H =∼ M as symmetric spaces. Any connected symmetric space M admits a “minimal” realization as a homogeneous symmetric space of its transvection group. The latter is defined as the subgroup of the automorphism group of M generated by all the products of an even number of symmetries. It is a finite dimensional Lie group which is transitive on M, stabilized by the conjugation (in the automorphism group) by so, for any o ∈ M, and minimal for these properties. A symmetric space is said to be semisimple (respectively, solvable) if its transvection group is semisimple (respectively, solvable). STRONGLY EXPONENTIAL SYMMETRIC SPACES 3

Every Lie group G is a symmetric space when endowed with the symmetries −1 sgl = gl g. It is usually realized by (G × G, Diag(G), σG) with σG(g, l) = (l, g), where G × G is in general larger than the transvection group of G. Given an involution σ on G, two canonical symmetric subspaces of G are deﬁned by P = −1 −1 {g ∈ G | σ(g) = g } and Gσ = {gσ(g) | g ∈ G}. When G is connected, Gσ is the connected component of the identity of P , and is isomorphic to G/Gσ through σ σ −1 the quadratic representation Q : G/G → Gσ : gG → gσ(g) . A symmetric subspace N of a symmetric space M =∼ G/H is said to be normal if there exists a σ-invariant normal subgroup R ⊂ G such that N =∼ R/(R ∩ H) through the isomorphism M =∼ G/H. Then M is said to be a semi-direct product

M1 n M2 of two symmetric subspaces M1 and M2 if M is, as a smooth manifold, a direct product M1 × M2, and if M2 is normal in M. On the tangent space m at a given point of a symmetric space M, there is a structure of Lie triple system [·, ·, ·]: m3 → m, generalizing that of Lie algebra for Lie groups. This trilinear bracket satisﬁes the properties:

(Lts1) [X,X,Y ] = 0, (Lts2) [X,Y,Z] + [Y,Z,X] + [Z,X,Y ] = 0, (Lts3) [X,Y, [U, V, W ]] = [[X,Y,U],V,W ] + [U, [X,Y,V ],W ] + [U, V, [X,Y,W ]].

Any Lie algebra g is a Lie triple system with bracket

(1) [X,Y,Z] = [[X,Y ],Z], and if G is a Lie group with Lie algebra g, this structure makes it the Lie triple system at e of G seen as a symmetric space. More generally, if σ is an involution on g and if g = m ⊕ h is the corresponding decomposition into (−1) and (+1)-eigenspaces of σ, then the bracket (1) on g restricts to m as a Lie triple system, and any Lie triple system is of that kind. The Lie triple system at the base point eH of a symmetric space G/H coming from a triple (G, H, σ) is given by the bracket (1) on the (−1)-eigenspace m of the involution dσe on Lie(G). Just as any connected symmetric space can be realized as a homogeneous symmetric space, any Lie triple system can be realized as the (−1)-eigenspace of a Lie algebra involution. The minimal such realization is called the standard embedding and is deﬁned (as a vector space) by g = m ⊕ [m, m] where [m, m] denotes the span of the operators LX,Y = [X,Y, ·] for all X,Y ∈ m. Conditions (Lts1)–(Lts3) show that the following equations deﬁne a Lie bracket on g:

[X,Y ] = LX,Y ,

[LX,Y ,Z] = LX,Y Z,

[LX,Y ,LU,V ] = LLX,Y U,V + LU,LX,Y V , 4 YANNICK VOGLAIRE for all X, Y, Z, U, V ∈ m. The involution σ on g is − Idm ⊕ Id[m,m] and one sees that (g, σ) is the infinitesimal data at the identity of the tranvection group. On any symmetric space, there is a unique affine connection for which all the symmetries are affine transformations. This connection is complete, torsion-free, and its curvature is parallel. The curvature R at a point x on a symmetric space is related to its Lie triple system at that point by R(X,Y )Z = −[X,Y,Z]. The exponential map at x is denoted Expx : TxM → M. On a symmetric space G/H with Lie triple system m ⊂ g at eH, the set exp m ⊂ G is an open submanifold of

P , so that exp m ⊂ Gσ ⊂ P are open inclusions. On a symmetric space with a fixed base point o, there is a product ⊥ defined ∼ by x ⊥ y = sxsoy. If M = G/H, one proves that s X so = exp X ∈ G for all Expo 2 X X ∈ m, so that the product satisfies Expo 2 ⊥ y = exp(X)y. A point z is said to be a midpoint of x and y if szx = y. Our first main result relates the uniqueness of midpoints to the exponentiality of a symmetric space.

Theorem 1.1. Let M be a connected symmetric space and (G, H, σ) be any symmetric triple realizing it. Denote [·, ·, ·] its Lie triple system at some base point o, and m the (−1)-eigenspace of σ in Lie(G). Then the following conditions are equivalent: (1) any two points in M have at most one midpoint; (2) the exponential map at o is injective; (3) the exponential map at o is a global diﬀeomorphism;

(4) M is simply connected and no operator Y 7→ [Y,X,X], for X ∈ ToM, has strictly negative eigenvalues; (5) M is simply connected and no operator ad X, for X ∈ m, has purely imaginary eigenvalues.

In particular, if every two points have at most one midpoint, then the space is exponential so that they actually have exactly one midpoint, and the map γ : M × M → M sending two points to their unique midpoint is smooth. By considering the symmetric space M = (G × G)/G associated to a connected Lie group G, in which case [X,Y,Z] = [[X,Y ],Z] for all X,Y,Z ∈ Lie(G), we recover the Dixmier–Saito theorem about the exponential map of solvable Lie groups. A symmetric space satisfying any of the conditions of Theorem 1.1 is said to be exponential. Examples (Ex5) and (Ex6) represent the two isomorphism classes of non-ﬂat solvable two-dimensional symmetric spaces, and are realized by (G, σ) where G is respectively SO(1, 1) n R2 and the universal cover of SO(2) n R2, and σ is the lift to G of the Lie algebra automorphism − Id ×σ2, with σ2(x, y) = (y, x). It is easily seen that the ﬁrst is exponential while the second is not. STRONGLY EXPONENTIAL SYMMETRIC SPACES 5

Theorem 1.2. If M is an exponential symmetric space, G is its transvection group σ and σ the conjugation by so for some o ∈ M, then G is connected, M is realized by (G, Gσ, σ), and (1) M is semisimple if and only if it is a Riemannian symmetric space of the noncompact type, (2) M is solvable if and only if G is an exponential Lie group,

(3) M is a semi-direct product M1 n M2 with M1 and M2 respectively of the X type (1) and (2), and the map m1 × m2 → M :(X,Y ) 7→ Expo 2 ⊥ Expo Y is a diﬀeomorphism.

In the solvable case, we can further characterize exponentiality in terms of (Ex6).

Theorem 1.3. Let M be a solvable, connected and simply connected symmetric space. The conditions of Theorem 1.1 are further equivalent to (6) no factor space of M has a subspace isomorphic to (Ex6).

We now turn to our last theorem. In an exponential symmetric space, a triple of points (x, y, z) is called a triangle, since any two points are joined by a unique geodesic. A double triangle of (x, y, z) is a triangle (a, b, c) whose midpoints of the edges are x, y, z. More precisely, sxc = a, sya = b and szb = c, so that szsya = sxa.

Theorem 1.4. Let M be an exponential symmetric space with transvection group G. Then every triangle has a unique double triangle if and only if M is solvable. In that case, for any g ∈ G and any z ∈ M, there exists a unique x ∈ M such that g · x = szx.

Since any element of the transvection group is a product of an even number of symmetries, the second statement is actually a generalization of the ﬁrst statement from triangles to n-gons, with odd n. This is to be related to the fact that in the aﬃne space Rn, any vector v (representing a translation, and thus a general transvection) can be “placed by its midpoint” anywhere: given a point z, the 1 translation of x = z − 2 v by v is the same as the image of x by the symmetry at z. This is not always possible in the hyperbolic plane for example, where if g is too large, there are points at which g cannot be “placed by its midpoint”. This idea and the notion of double triangle appear in the study of generating functions for canonical tranformations on symplectic symmetric spaces [20, 13, 19]. The paper is organized as follows. In Section2, we prove Theorems 1.1, 1.2 and 1.3, and in Section3, we prove Theorem 1.4.

2. Midpoints and exponential spaces

In this section, M will always denote a connected symmetric space, o ∈ M a ﬁxed base point, and m the Lie triple system of M at o. For a symmetric triple 6 YANNICK VOGLAIRE

(G, H, σ) realizing M, we will denote g the Lie algebra of G, and g = h ⊕ m its decomposition into (±1)-eigenspaces for σ, identifying the (−1)-eigenspace with the Lie triple system. A symmetric pair (G, σ) (respectively, a symmetric triple (G, H, σ)) realizing M will be called a transvection pair (respectively, a transvection triple) if G is the transvection group of M, and σ is the conjugation by so.

Deﬁnition 2.1. A point z ∈ M is said to be a midpoint for x, y ∈ M if szx = y, and is said to be a square root of x if z is a midpoint for x and o.

In order to motivate the deﬁnition of local exponentiality, let us recall Helgason’s formula and the well-known fact that it implies.

Proposition 2.2 (Helgason’s formula [7, Theorem 4.1]). Let (G, H, σ) be a symmetric triple, and for g ∈ G denote τ(g) the map hH 7→ ghH from G/H onto itself.

The diﬀerential of the exponential map Expo from m to G/H is given by sinh ad X d(Exp ) = d(τ(exp X)) ◦ . o X eH ad X

Lemma 2.3. Let (G, H, σ) be a symmetric triple realizing M. Then Expo is a local diﬀeomorphism at X ∈ m if and only if the set of eigenvalues of ad X only meets iπZ at 0.

Proof. By Proposition 2.2, Expo is a local diﬀeomorphism at X ∈ m if and only if sinh(ad X) sinh λ det ad X =6 0, a condition which is equivalent to λ =6 0 for every eigenvalue λ of ad X, which in turn amounts to the statement. Lemma 2.4. Let (G, H, σ) be a symmetric triple realizing M, and let X ∈ m. Then a complex number λ is an eigenvalue of ad X if and only if its square is an eigenvalue of Y ∈ m 7→ [Y,X,X].

2 Proof. Seeing m as a subset of g, we have [ · ,X,X] = [[ · ,X],X] = (ad X) |m. Set m = dim m and h = dim h. Choosing a basis of g = m ⊕ h, we may write 0 A ad X = B 0 where A and B are m × h and h × m matrices respectively, and we 2 AB 0 have (ad X) = 0 BA . By Sylvester’s determinant formula, we get det(AB − m−h 2 λIm) = (−λ) det(BA − λIk), thus the eigenvalues of AB = (ad X) |m and 2 BA = (ad X) |h are identical, up to |m − h| zeros, which yields the conclusion. Deﬁnition 2.5. The symmetric space M is said to be locally exponential if no operator m → m : Y 7→ [Y,X,X], X ∈ m, has strictly negative eigenvalues, and exponential if the exponential map Expo : ToM → M at o is a global diﬀeomorphism.

By translation, in an exponential symmetric space, every exponential map Expx, x ∈ M, is a diﬀeomorphism. Also, it is clear that a symmetric space is exponential if and only if any symmetric triple (G, H, σ) realizing it is exponential in the sense that the map exp : m → G/H is a global diﬀeomorphism. STRONGLY EXPONENTIAL SYMMETRIC SPACES 7

Lemma 2.6. If every two points have at most one midpoint, then the exponential map Expo is injective. X Proof. Assume X,Y ∈ ToM are such that Expo X = Expo Y . Then, Expo 2 and Y Expo 2 are both square roots of Expo X, so that they coincide by assumption. By X Y induction, we obtain Expo n = Expo n for all positive integers n. Since Expo is injective around 0, we have X = Y . Recall that the automorphisms of a symmetric space are the affine automorphisms of its canonical affine structure. In particular, the exponential map at some point o is equivariant under the actions of the subgroup of automorphisms fixing o.

Lemma 2.7. If the exponential map Expo is injective, then it is an immersion, and M is diﬀeomorphic to a Euclidean space. In particular, M is simply connected and locally exponential.

Proof. Let (G, H, σ) be a triple realizing M. Assume that Expo is not an immersion at X ∈ m, so that ad X has an eigenvalue kπi for some k ∈ Z∗. We will show that there is a curve through X in m on which Expo is constant, so that Expo is not injective. Let 0 6= Y ∈ gC be such that [X,Y ] = kπiY . Spelling out the eigenvalue equation, one sees that Y = (A1 + iB1) + i(A2 + iB2) with Aj ∈ m,

Bj ∈ h, [X,Aj] = kπBj, and [X,Bj] = −kπAj, for j = 1, 2. We can assume

t ad B1 that A1 6= 0 and deﬁne X(t) = e X for all t ∈ R. Then, on the one hand, [B1,X] = kπA1 6= 0 so X(t) is not constant, and on the other hand sinh(ad X(t)) sinh(ad X) X˙ (t) = et ad B1 e−t ad B1 X˙ (t) ad X(t) ad X sinh(ad X) = et ad B1 [B ,X] ad X 1 t ad B1 = e sin(kπ)A1 = 0,

d so that dt Expo(X(t)) = 0 for all t ∈ R by Helgason’s formula. Hence t 7→ X(t) is the curve we were looking for. For the second assertion, in [8, Theorem 1] and [9, Chapter IV, Theorem 3.5], it is shown that any connected symmetric space G/H is smoothly fibered over a (connected) compact Riemannian symmetric subspace N = K/(K ∩ H) with fibers diffeomorphic to a Euclidean space. But no non-discrete compact symmetric space has injective exponential maps, so N must be reduced to a point. The last assertion follows now from the first two assertions and from Lemma 2.3. Lemma 2.8. Let M be exponential. Then every two points have a unique midpoint.

−1 Proof. Let x, y ∈ M, and c : R → M be the geodesic c(t) = Expx(t Expx (y)). It is easily checked that z = c(1/2) is a midpoint for x and y, and that it is unique. 8 YANNICK VOGLAIRE

To prove that simple connectedness and local exponentiality together imply exponentiality, we ﬁrst address the solvable and semisimple cases.

Lemma 2.9. Let M be simply connected, and (G, H, σ) be a transvection triple realizing it. Denote G˜ the universal cover of G, and σ˜ : G˜ → G˜ the lift of σ to G˜. ∼ σ ∼ σ˜ σ˜ Then M = G/(G )0 = G/˜ G˜ and G˜ is connected.

σ σ Proof. We know that (G )0 ⊂ H ⊂ G . But since G is connected and M is simply ∼ σ connected, H must be connected, which yields M = G/(G )0. Denote π : G˜ → G the covering map. Recall that the fixed point subgroup of an involution of a simply connected Lie group is always connected [9, Chapter IV, Theorem 3.4], so G˜σ˜ is connected. On the other hand, we have M =∼ G/π˜ −1(H). σ˜ −1 But (G˜ )0 is an open subgroup of π (H) and, by the same argument as in the −1 −1 ˜σ˜ ˜σ˜ foregoing paragraph, π (H) is connected. Hence π (H) = (G )0 = G . Proposition 2.10 ([1, Proposition 2.1]). Let (G, σ) be a symmetric pair, with G exponential. Then, (1) the exponential map is a diffeomorphism from h onto Gσ; (2) the exponential map is a diffeomorphism from m onto P ; (3) the multiplication m is a diffeomorphism from P × Gσ onto G.

Proposition 2.11 ([9, Corollary, p. 171]). Let M be a simply connected symmetric space with transvection group G, and let G˜ be its universal cover. Then G =∼ G/Z˜ (G˜)σ where Z(G˜)σ = {z ∈ Z(G˜) | σ(g) = g}.

Proposition 2.12. Let M be solvable. If M is connected, simply connected, and locally exponential, then M and its transvection group G are exponential.

Proof. With the notation of Lemma 2.9, we have M =∼ G/˜ G˜σ. Since g is solvable and is the transvection Lie algebra, we see that h = [m, m] is contained in the nilradical of g so that for all X ∈ h, ad X only has null eigenvalues. By Lie’s theorem, the complexiﬁcation of g may be represented by upper triangular matrices, and the representatives of h must then be strictly upper triangular. This shows that the eigenvalues of an element X = Xh + Xm ∈ g are equal to those of Xm ∈ m, which cannot be purely imaginary. Hence G˜ is exponential by the Dixmier–Saito Theorem, and consequently G/˜ G˜σ˜ is exponential by Proposition 2.10. We now show that G = G˜. By Proposition 2.11, all we need to see is that Z(G˜) ∩ G˜σ˜ is reduced to a point. Let z ∈ Z(G˜) ∩ Gσ. Since G˜ is exponential, there exists a unique Z ∈ g such that z = eZ . Since z is in the center of G˜, we have

Z Z −1 Adg Z e = ge g = e , so that by exponentiality, Z = Adg Z for all g ∈ G˜, and Z is in the center of g. But z is also in G˜σ˜ , so that Z ∈ h. Since g is the transvection algebra, the action of h on m is eﬀective, and the only central element in h is 0. Hence z = e. STRONGLY EXPONENTIAL SYMMETRIC SPACES 9

Proposition 2.13. Let M be semisimple. If M is connected, simply connected and locally exponential, then it is a Riemannian symmetric space of the noncompact type. In particular, it is exponential.

Proof. Let (G, H, σ) be a transvection triple realizing M where, by hypothesis, G is semisimple. Let θ be a Cartan involution commuting with σ ([2], [9, p. 153, Theorem 2.1]), and B be the Killing form of g. We have the decompositions in eigenspaces

g = h ⊕ m for σ = k ⊕ p for θ

= hk ⊕ hp ⊕ mk ⊕ mp, where the spaces of the last line are the obvious intersections of those of the ﬁrst two lines. Now if 0 6= Z ∈ mk, then Z is in k so ad Z is a skewsymmetric matrix for the scalar product Bθ(X,Y ) = −B(X, θY ). Since the adjoint representation is faithful, ad Z has non-zero purely imaginary eigenvalues, contradicting local exponentiality.

Hence mk = {0}. But then h = [m, m] = [mp, mp] = hk, so that hp = 0. We have thus shown that σ = θ. Since σ is a Cartan involution on G and G is connected, the global polar decomposition for Riemannian symmetric spaces of the noncompact type [7, Chapter VI, Theorem 1.1] implies that M is exponential. Using these two special cases, we may now prove the general result by induction on the dimension of the radical. The argument is inspired from [15].

Theorem 2.14. If M is simply connected and locally exponential, then it is exponential. Moreover, it is a semidirect product M1 nM2 of a semisimple and a solvable X exponential subspaces, and the map m1 × m2 → M :(X,Y ) 7→ Expo 2 ⊥ Expo Y is a diﬀeomorphism.

Proof. Let G be the universal cover of the transvection group of M, so that M =∼ G/Gσ, and let R be the radical of G. If R is 0-dimensional, the theorem reduces to Proposition 2.13, so let us assume the contrary. Since G is simply connected, there exists a global Levi decomposition G = S n R —with S semisimple and R the non-trivial radical, both simply connected— which is compatible with σ in the sense that the latter stabilizes both S and R. Let N be a minimal σ-invariant normal subgroup of G. It is contained in R, and is connected since otherwise its connected component containing e would be preserved by σ and, G being connected, by conjugation by G. Similarly, N is Abelian, since otherwise the commutant would be a smaller connected σ-invariant normal subgroup. And since it is a connected subgroup of a simply connected solvable Lie group, N 10 YANNICK VOGLAIRE is also closed and simply connected [10]. Moreover, as the fixed point subgroup of an involution of a simply connected Lie group, N σ is connected, so that N/N σ is simply connected. Let Gˆ = G/N, let π : G → Gˆ be the natural projection, σˆ the induced involution and π0 : G/Gσ → G/ˆ Gˆσˆ the natural morphism. Since G/Gσ is connected and simply connected, and N/N σ is connected, G/ˆ Gˆσˆ is connected and simply connected. Moreover, the eigenvalues of ad mˆ are to be found amongst those of ad m, so that G/ˆ Gˆσˆ is locally exponential and, by the inductive assumption, is exponential. σ 0 Let x ∈ G/G . Then there exists a unique X ∈ mˆ such that ExpeGˆσˆ X = π (x). Denote F ⊂ G the complete preimage through π of {exp tX | t ∈ R}. It is a σ- 0 0 0 invariant subgroup isomorphic to exp hX inN for some X ∈ m such that π∗X = X. Since N/N σ is connected and simply connected, F/F σ = F/N σ also is, and thus the latter inherits local exponentiality from G/Gσ. Since F is solvable, F/F σ is then exponential by Proposition 2.12 so that, denoting f the Lie algebra of F , there 00 00 exists a unique X ∈ f ∩ m such that ExpeGσ X = x. σ We thus have a bijection ExpeGσ between m and G/G , and the local exponentiality of M shows that it is everywhere a local diffeomorphism, hence a global diffeomorphism. We now turn to the second assertion. Consider the natural map p : G/Gσ → σ ∼ σ σ σ (G/R)/(G/R) = S/S , and set M1 = S/S and M2 = R/R . An element of M1 σ may be uniquely written as Expo X = exp(X)S for some X ∈ m1. The complete σ σ inverse image of Expo X by p is {exp(X).rG | r ∈ R} = {exp(X). exp(Y )G | X X Y ∈ m2} = {Expo( 2 ) ⊥ Expo(Y ) | Y ∈ m2}, and the map (X,Y ) 7→ Expo( 2 ) ⊥ Expo(Y ) is clearly a diffeomorphism.

Proposition 2.15. Let M be diﬀeomorphic to Rn. If (G, σ) is a transvection pair, then Gσ is connected. Moreover, every connected symmetric subspace is diﬀeomorphic to Rn.

Proof. Assume Gσ is not connected. Since the fixed points subgroup of an involution of a connected Lie group always has a finite number of connected components [9, Chapter IV, Theorem 3.4], G/Gσ is then at the same time a symmetric space, and thus a manifold, and a quotient of Rn by the (effective) action of the finite σ σ group G /(G )0, hence an orbifold, a contradiction. For the second assertion, by [8, Theorem 1], the maximal compact subspace of M is trivial, hence any subspace of M has the same property, and the same theorem yields the conclusion.

We can now assemble the above results to prove Theorems 1.1 and 1.2. STRONGLY EXPONENTIAL SYMMETRIC SPACES 11

Proof of Theorem 1.1. The assertion (1)⇒(2) follows from Lemma 2.6, (2)⇒(4) follows from Lemma 2.7, (4)⇔(5) is proved in Lemma 2.4, (4)⇒(3) is Theorem 2.14, and ﬁnally (3)⇒(1) follows from Lemma 2.8.

Proof of Theorem 1.2. If M is exponential, (1) follows from Proposition 2.13, (2) follows from Proposition 2.12, and (3) follows from Theorem 2.14.

Example 2.16. The (canonical structure on the) tangent bundle of a locally exponential Lie triple system m, deﬁned by T m := m ⊕ m with

[(x, u), (y, v), (z, w)] = ([x, y, z], [u, y, z] + [x, v, z] + [x, y, w]), is also locally exponential. Indeed, any operator [(·, ·), (y, v), (y, v)] has the form [·,y,y] 0 [·,v,y]+[·,y,v][·,y,y] , and thus has the same eigenvalues as [·, y, y]. This proves that the tangent bundle of a Riemannian symmetric space of the noncompact type is a non-semisimple, non-solvable, exponential symmetric space. C

Example 2.17. A midpoint map is a smooth map γ : M × M → M such that sγ(x,y)x = y for all x, y ∈ M. There exist many ﬁnite symmetric spaces with midpoint map (those spaces were called ﬁnite symsets and studied by Nobusawa and his collaborators [11]), the simplest non-trivial example being the space M =

{x1, x2, x3} where the symmetry at a point interchanges the remaining two points, and the midpoint of two points yields the remaining point. However, none of these finite spaces (apart from the 1-point space) embeds in any connected symmetric space with midpoint map. Indeed, assume we have such an embedding and fix two points o, x0 ∈ M. Then the set {xn}n defined by xn = γ(o, xn−1) is finite. But one sees that it lies entirely on one single geodesic, so that the successive midpoints with o should actually converge to o, contradicting the assumption. C

Example 2.18. In [3], in order to build non-formal deformation quantizations, a class of symmetric spaces is introduced as the first step of a potential inductive construction of all solvable symplectic symmetric spaces. These (connected and simply connected) elementary solvable symplectic spaces are realized by triples (G, H, σ) where the Lie algebra of G is a semi-direct product g = a n b of two abelian Lie subalgebras stabilized by σ, such that there exists ξ ∈ g∗ such that the 2-coboundary Ω: g × g → R defined by Ω(X,Y ) = − hξ, [X,Y ]i restricts as a non-degenerate two-form on the (−1)-eigenspace m of σ in g. This implies that b = l ⊕ h with l ⊂ m, and that a and l are two lagrangians in involution for Ω|m×m. A quantization is built in [3] for all elementary solvable symplectic spaces which admit a midpoint map. It is then proved that such a map exists if and only if the twisting map φ : a → a, defined by Ω(φ(A),L) = hξ, sinh(ad A)Li for all A ∈ a and L ∈ l, is a global diffeomorphism. It is easily seen that, if φ is a diffeomorphism then, 12 YANNICK VOGLAIRE

2 for all A ∈ a, cosh(ad A): l → l is invertible so that, a being abelian, (ad A) |m has no negative eigenvalues. Theorem 1.1 implies that the converse is also true, giving a simple characterization of the spaces for which the constructions in [3] apply. C Corollary 2.19. Let M be exponential. Then (1) a symmetric subspace N is exponential if and only if midpoints of points in N are in N, if and only if N is connected; (2) the intersection of any two exponential symmetric subspaces is empty or exponential; (3) the quotient M/N by a closed normal symmetric subspace is exponential if and only if N is exponential.

Proof. (1) If N is exponential, then it obviously satisﬁes the other two assertions. Assume that midpoints of points in N are in N. Let x, y ∈ N. There is a unique geodesic c : [0, 1] → M joining x and y in M. By taking recursively the midpoints of x and y and of their successive midpoints, we obtain that n m the dyadic geodesic c( 2m | m ∈ N, 0 ≤ n ≤ 2 ) is in N, and by density, that the whole geodesic c([0, 1]) is in N. Hence, N is exponential. Assume now that it is connected. Then it is simply connected by Propo- sition 2.15, and it is locally exponential since M is. Hence it is exponential by Theorem 1.1.

(2) Assume the intersection N = N1 ∩ N2 is not empty. Then any two points in N are joined by a unique geodesic in M which, by assumption, lies in

both N1 and N2. (3) Every eigenvalue of y 7→ [y, x, x] for x ∈ m/n is also an eigenvalue of Y 7→ [Y,X,X] for some X ∈ m, so M/N is locally exponential. On the other hand, it is clear that M/N is simply connected if and only if N is connected. We conclude this section by giving a characterization of the exponentiality of solvable symmetric spaces in terms of (Ex6). To this end, we ﬁrst adapt the notions of Jordan–Hölder series and roots to modules of (solvable) Lie triple systems. If m is a Lie triple system, an m-module is a vector space V together with a structure of Lie triple system on m ⊕ V such that m is a subsystem and V is an abelian ideal (an ideal is abelian if the triple bracket is zero whenever two of its arguments are in that ideal). An m-module is simple if it has no non-zero proper submodule. Let m be a solvable Lie triple system over C. Then any simple m-module over C is of dimension 1. Indeed, let V be such a module and let g = m + [m, m] (respectively, gV = (m + V ) + [m + V, m + V ] = g + (V + [m,V ])) be the standard embedding of m

(respectively, of m+V ). Let h be a minimal ideal of gV contained in the abelian ideal STRONGLY EXPONENTIAL SYMMETRIC SPACES 13

V + [m,V ]. It is of dimension 1. And h + σh = V + [m,V ] since h + σh ⊂ V + [m,V ] by σ-invariance and h + σh ⊃ V + [m,V ] since otherwise (m + V ) ∩ (h + σh) would be a proper submodule of V . So either h = σh or h ∩ σh = {0}. In both cases, dim V = 1. Let m be a solvable Lie triple system over R. Then any simple m-module over R is of dimension 1 or 2. Indeed, let V , g, gV and h be as in the foregoing paragraph. Then h is of dimension 1 or 2, and either h = σh or h ∩ σh = {0}, which yields the following four possibilities:

(S1) h = σh and dim h = 1: dim V = 1 and V is central in m + V ,

(S2) h = σh and dim h = 2: dim V = 1 and for all X ∈ m, [·,X,X]|V has real non-positive eigenvalue,

(S3) h ∩ σh = {0} and dim h = 1: dim V = 1 and for all X ∈ m, [·,X,X]|V has real non-negative eigenvalue,

(S4) h ∩ σh = {0} and dim h = 2: dim V = 2 and for all X ∈ m, [·,X,X]|V has null or non-real eigenvalues.

In the last three cases, there exists at least a generator X ∈ m such that [·,X,X]|V has non-zero eigenvalues.

A Jordan–Hölder series for an m-module V is a decreasing sequence V = V0 ⊃

V1 ⊃ · · · ⊃ Vp = {0} of submodules of V such that each quotient Vk/Vk+1 is simple. Every module for a solvable Lie triple system m has a Jordan–Hölder series. Indeed, let m be solvable and let V be an m-module of dimension n. Since m is solvable, m + V also is, and so is its standard embedding gV . Let h be a minimal σ-invariant ideal of gV contained in V + [m,V ]. Then W = (m + V ) ∩ h is a simple ideal of m + V contained in V , i.e., a simple submodule of V . Assume by induction that any m-module of dimension not greater than n − 1 has a Jordan–Hölder series, and let 0 0 0 0 0 V0 = V ⊃ V1 ⊃ · · · ⊃ Vn−1 = {0} be such a series for V = V/W . If π : V → V/W −1 0 denotes the canonical projection, then setting Vk = π (Vk) for k = 0, . . . , n − 2, and Vn−1 = W , Vn = {0} defines a Jordan–Hölder series for V . The Jordan–Hölder theorem holds for m-modules since they can be viewed as groups with operators [4, pp. 30–31, 41] (V, m × m) with V seen as an abelian group, the action being (m × m) × V → V : ((X,Y ), v) 7→ [v, X, Y ]. Given a solvable Lie triple system m over R, an m-module V of dimension n and a Jordan–Hölder series V = V0 ⊃ V1 ⊃ · · · ⊃ Vp = {0} for V , the weights of V are defined as follows. Let mC, VC and Vk,C be the complexifications of m, V and Vk, k = 0, . . . , p respectively. If, for some k, Vk,C/Vk+1,C is not simple as an mC-module 0 (this will only happen in the case (S4) above), insert a submodule Vk between Vk,C 0 0 and Vk+1,C such that Vk,C/Vk and Vk/Vk+1,C are simple. Doing this for all k yields ˜ a Jordan–Hölder series Vl, l = 0, . . . , n − 1 for VC. For each l, by restricting to 14 YANNICK VOGLAIRE

˜ ˜ ˜ Vl and passing to the quotient Vl/Vl+1, the map mC → End(VC): X 7→ [·,X,X] ˜ ˜ ˜ ∼ ˜ deﬁnes a quadratic form φl : mC → End(Vl/Vl+1) = C. Restricting the forms φl to m ⊂ mC yields the weights φl : m → C of V as an m-module. Up to the order, they are independent of the choice of Jordan–Hölder series for V and of the choice of 0 submodules Vk. When the m-module is m itself with the module structure described in Example 2.16, the weights are called the roots of m.

Proof of Theorem 1.3. If a factor space of M has a subspace isomorphic to (Ex6) then, by Corollary 2.19, M is not exponential.

Conversely, assume that M is not exponential. Let m = m0 ⊃ m1 ⊃ · · · ⊃ mp = {0} be a Jordan–Hölder series for m and φl, l = 0,..., dim m − 1 be the corresponding roots as constructed above. Then, since m is not locally exponential, there exists a root φl of m corresponding to a simple quotient mk/mk+1 of the kind

(S2) above. Let Xˆ ∈ m be such that φl(Xˆ) = −1, let Yˆ ∈ mk \ mk+1, and let X,Y be their respective images in m/mk+1. Then, in m/mk+1, we have [Y,X,X] = −Y , and moreover [Y,X,Y ] = 0 since mk/mk+1 + [m, mk/mk+1] is abelian. Hence X and Y span a Lie triple system isomorphic to that of (Ex6), which by Proposition 2.15 implies the result.

3. Double triangles and solvable spaces

In this section, M will denote an exponential symmetric space with base point o, and γ : M × M → M its midpoint map: the smooth map sending two points to their unique midpoint.

Deﬁnition 3.1 ([12]). The map γ3 : M × M × M → M × M × M is deﬁned by

γ3(x, y, z) = (γ(z, x), γ(x, y), γ(y, z)) for all x, y, z ∈ M. A triple (a, b, c) such that γ3(a, b, c) = (x, y, z) is called a double triangle of (x, y, z).

Example 3.2. It was shown in [18, 13, 14] that there exist triangles (x, y, z) in the hyperbolic plane that have no double triangle, so that γ3 is not a diﬀeomorphism. If a given triangle has a double triangle, however, the latter is unique. Seeing the 2 2 hyperbolic plane as the x3 > 0 sheet of the two-sheeted hyperboloid (x3) − (x1) − 2 3 (x2) = 1 in R , the condition [18] for the existence of a double triangle of a triangle (x, y, z) is that det(xyz)2 < 1, where xyz denotes the 3 × 3 matrix whose columns are the vectors x, y and z. C

Theorem 3.3. If M is solvable, then γ3 is a diﬀeomorphism. Moreover, for any odd natural number n, every n-gon has a unique double in the sense that the obvious n n generalization γn : M → M is a diﬀeomorphism. STRONGLY EXPONENTIAL SYMMETRIC SPACES 15

Proof. Let (G, σ) be a transvection pair, so that M =∼ G/Gσ and G is exponential. We will prove the result by induction on the dimension of G. If G is 1-dimensional, then M is isomorphic to R with the usual symmetries sxy = 2x − y. Hence, δ3(x, y, z) = (y − z + x, z − x + y, x − y + z) is a smooth inverse of γ3. Assume now that the dimension of G is greater than 1, and let N be a minimal σ-invariant normal subgroup of G. As in the proof of Theorem 2.14, we can see that N is a closed abelian connected and simply connected normal subgroup, and by Proposition 2.15, N/N σ is also simply connected. Set Gˆ = G/N. Since G/ˆ Gˆσˆ is contractible, the canonical morphism π : G/Gσ → G/ˆ Gˆσˆ (which is a ﬁbre bundle) admits a smooth global section s [17, pp. 25, 55–56]. Moreover, since G/Gσ is exponential, there is a smooth global section t of the σ −1 canonical projection ρ : G → G/G with values in Gσ = {gσ(g) | g ∈ G} ⊂ G (it X is given by t(Expo X) = Q(Expo 2 ) where Q is the quadratic representation deﬁned in the introduction). To summarize, we have the following commutative diagram, where the arrows on the two upper lines are group morphisms, and those on the lower line are symmetric space morphisms: N σ / Gσ / Gσ/N σ

N / G / G/N U ρ t s N/N σ G/Gσ r G/ˆ Gˆσˆ / π / Thus the composition t ◦ s gives a map G/ˆ Gˆσˆ → G, and for each x ∈ G/ˆ Gˆσˆ the set {t(s(x)).n | n ∈ N} is mapped through ρ onto the ﬁber π−1(x). Moreover, since N is abelian connected and simply connected, it is exponential so that by Proposition 2.10 there exists a smooth section r of N → N/N σ with values in the Lie subgroup N 0 = {n ∈ N | σ(n) = n−1} of N. With the help of the above constructions, we may uniquely represent any element x ∈ G/Gσ as

x = gnGσ where g = t(s(π(x))) and n = r(g−1x) ∈ N 0.(2)

σ Let now x1, x2, x3 ∈ G/G . By the induction assumption, there exists a unique ˆ ˆσˆ xˆ ∈ G/G such that sπ(x3)sπ(x2)sπ(x1)xˆ = x,ˆ and xˆ is a smooth function of σ (π(x1), π(x2), π(x3)). We will build a unique x ∈ G/G depending smoothly on 16 YANNICK VOGLAIRE

(x1, x2, x3) and such that

(3) sx3 sx2 sx1 x = x.

If x is a solution of (3), it must be such that π(x) = xˆ by uniqueness of xˆ. Hence, writing x = gnGσ with g = t(s(xˆ)) and n ∈ N 0, solving equation (3) amounts to σ solving an equation on n. Writing xi = giG for some gi ∈ G, we have

σ −1 −1 −1 σ −1 σ (4) sx3 sx2 sx1 (gnG ) = g3σ(g3 g2)g2 g1σ(g1 g)σ(n)G =gn ˜ G ,

−1 −1 −1 where g˜ = g3σ(g3 g2)g2 g1σ(g1 g). But by (2) and the deﬁnitions of g and xˆ, we have gG˜ σ = gnG˜ σ for a unique n˜ ∈ N 0. Hence there further exists a unique k ∈ Gσ such that g˜ = gnk˜ , and the right hand side in equation (4) is equal to −1 σ −1 0 σ 0 g Ck(n )nG˜ , where Ck(n ) ∈ N since k ∈ G and n ∈ N , and where we used 0 σ −1 the fact that N is abelian. This will be equal to gnG if and only if n = Ck(n )n˜.

Replacing the n’s by their log, we get (Id + Adk) (n) = n.˜ Now G is exponential X so k = e with X ∈ k nilpotent, so that Id + Adk is invertible and its inverse is −1 polynomial in ad X. We thus have a unique solution n = (Id + Adk) n˜ which is σ smooth in (x1, x2, x3), and consequently a unique solution x = t(s(xˆ))nG of the equation sx3 sx2 sx1 x = x with the same property. Hence γ3 is invertible and has a smooth inverse. As γ3 itself is smooth, we have a diﬀeomorphism. For the second assertion of the theorem, just notice that the argument above only relied on the fact that there was an odd number of symmetries in equation (3).

Proposition 3.4. If γ3 is a diﬀeomorphism and if N is a connected symmetric 3 3 subspace of M, then γ3|N 3 is a diﬀeomorphism N → N .

Proof. Since N is connected, by Corollary 2.19 (1), γ restricts to N as a midpoint 3 3 3 3 map, so that γ3(N ) ⊂ N . Obviously, γ3(N ) is open in N . But N is closed in 3 3 3 M, so γ3(N ) is closed in M and thus in N . Proof of Theorem 1.4. As in Lemma 2.9, let us realize M as G/˜ G˜σ, where G˜ is the universal cover of the transvection group of M. Now there exists a global Levi decomposition G˜ = S n R, with S and R stable under σ, S semisimple and R the radical. Thus S/Sσ is a semisimple symmetric subspace of M. 0 Let us prove that for any semisimple exponential symmetric space M , γ3 is not surjective so that, by Proposition 3.4, S/Sσ must be reduced to a point and M must be solvable. To realize M 0, by Theorem 1.1 it is suﬃcient to consider (G, θ) a Riemannian symmetric pair of the noncompact type. Then there exists a symmetric subspace of M 0 isomorphic to the hyperbolic plane [7, Chapter IX] for which, by

Example 3.2, γ3 is not surjective. Proposition 3.4 then implies that so is the case for M 0. The converse implication of the theorem is the content of Theorem 3.3. STRONGLY EXPONENTIAL SYMMETRIC SPACES 17

References

[1] Y. Benoist. Multiplicité un pour les espaces symétriques exponentiels. Mémoires de la Société Mathématique de France. Nouvelle Série, (15):1–37, 1984. Harmonic analysis on Lie groups and symmetric spaces (Kleebach, 1983). [2] M. Berger. Les espaces symétriques noncompacts. Annales Scientifiques de l’École Normale Supérieure. Troisième Série, 74:85–177, 1957. [3] P. Bieliavsky. Strict quantization of solvable symmetric spaces. The Journal of Symplectic Geometry, 1(2):269–320, 2002. [4] N. Bourbaki. Algebra I. Chapters 1–3. Elements of Mathematics (Berlin). Springer-Verlag, Berlin, 1998. [5] J. Dixmier. L’application exponentielle dans les groupes de Lie résolubles. Bulletin de la Société Mathématique de France, 85:113–121, 1957. [6] D. Djoković and K. H. Hofmann. The surjectivity question for the exponential function of real Lie groups: a status report. Journal of Lie Theory, 7(2):171–199, 1997. [7] S. Helgason. Differential geometry, Lie groups, and symmetric spaces, volume 80 of Pure and Applied Mathematics. Academic Press Inc. [Harcourt Brace Jovanovich Publishers], New York, 1978. [8] S. Koh. On affine symmetric spaces. Transactions of the American Mathematical Society, 119(2):291–309, Aug. 1965. [9] O. Loos. Symmetric spaces. I: General theory. W. A. Benjamin, Inc., New York-Amsterdam, 1969. [10] A. Malcev. On the theory of the Lie groups in the large. Matematicheskii Sbornik, 16(58):163–190, 1945. [11] N. Nobusawa. On symmetric structure of a finite set. Osaka Journal of Mathematics, 11:569–575, 1974. [12] Z. Qian. Groupoids, Midpoints and Quantizations. PhD thesis, University of California, Berkeley, 1997. [13] P. d. M. Rios and A. Ozorio de Almeida. A variational principle for actions on symmetric symplectic spaces. Journal of Geometry and Physics, 51(4):404–441, 2004. [14] P. d. M. Rios and G. M. Tuynman. Weyl quantization from geometric quantization. In Geometric methods in physics, volume 1079 of AIP Conference Proceedings, page 26–38. American Institute of Physics, Melville, NY, 2008. [15] P. K. Rozanov. On the exponentiality of affine symmetric spaces. Rossi˘ıskayaAkademiya Nauk. Funktsionalny˘ıAnaliz i ego Prilozheniya, 43(1):68–80, 2009. [16] M. Saito. Sur certains groupes de Lie résolubles. Scientific Papers of the College of Arts and Sciences. The University of Tokyo, 7:1–11, 1957. [17] N. Steenrod. The Topology of Fibre Bundles. Princeton Mathematical Series, vol. 14. Princeton University Press, Princeton, N. J., 1951. [18] G. M. Tuynman. Areas of spherical and hyperbolic triangles in terms of their midpoints. U. Lille 1-Math, 1999. Prepublication, http://arxiv.org/abs/1307.2567. [19] P. Ševera. Symplectic generating functions and moyal products. Letters in Mathematical Physics, 75(3):273–277, 2006. [20] A. Weinstein. Traces and triangles in symmetric symplectic spaces. In Symplectic geometry and quantization (Sanda and Yokohama, 1993), volume 179 of Contemporary Mathematics, page 261–270. American Mathematical Society, Providence, RI, 1994. 18 YANNICK VOGLAIRE

[21] M. Wüstner. A short survey on the surjectivity of exponential Lie groups. In Nuclear groups and Lie groups (Madrid, 1999), volume 24 of Research and Exposition in Mathematics, page 239–249. Heldermann, Lemgo, 2001.

Department of Mathematics, The Pennsylvania State University, 104 McAllister Bldg, University Park, 16802 PA, USA, E-mail address: [email protected]